Efficient Discrete Clustering With Anchor Graph.

Area of Science:

• Machine Learning
• Data Mining
• Graph Theory

Background:

• Spectral clustering (SC) is a powerful technique for graph learning but suffers from computational inefficiency due to eigenvalue decomposition (EVD).
• Large-scale datasets exacerbate issues of time consumption and information loss in traditional SC methods.
• Existing SC methods often require post-processing steps like binary label optimization, impacting overall efficiency.

Purpose of the Study:

• To propose a fast and efficient clustering method for large-scale data.
• To address the limitations of spectral clustering, including computational cost and information loss.
• To develop a method that avoids post-processing steps through direct discrete label optimization.

Main Methods:

• Efficient Discrete Clustering with Anchor Graph (EDCAG) method is proposed.
• Utilizes sparse anchors to accelerate graph construction and create a parameter-free similarity matrix.
• Employs an intraclass similarity maximization model between anchor-sample layers.
• Applies a fast coordinate rising (CR) algorithm for optimizing discrete labels of samples and anchors.

Main Results:

• EDCAG demonstrates significant improvements in speed compared to traditional spectral clustering.
• The method achieves competitive clustering performance, maintaining accuracy on large datasets.
• Circumvents the need for post-processing steps like binary label optimization.

Conclusions:

• EDCAG offers a rapid and effective solution for large-scale data clustering.
• The proposed method enhances the practicality of spectral clustering for real-world applications.
• Anchor graph-based approaches combined with coordinate rising optimization show promise for efficient graph learning.